PDS 69: Methods and statistics, Johan Friso Foyer, Floor 1, August 26, 2019, 4:30 PM - 5:30 PM
Background/Aim: Two-stage analyses are widely applied in environmental research to investigate short-term effects of environmental stressors. The standard design involves first-stage models to derive single association estimates from multiple locations, and their pooling in second-stage random-effects meta-analysis. However, this approach can be limited for modern epidemiological investigations. In this contribution, we propose an extended two-stage framework that allows the definition of more flexible multi-location designs for environmental studies.
Methods: The extended two-stage framework is based on newly developed meta-analytical models that can incorporate multiple hierarchical levels and repeated location-specific estimates. This allows consideration of additional levels of spatial clustering, the estimation of associations at different times or for different groups, and the analysis of effect modification by location-specific characteristics varying both in time and in space. Applications are facilitated by the implementation of the meta-analytical models in the new R package mixmeta.
Applications: The extended framework is illustrated in four applications using the Multi-Country Multi-City (MCC) database, including time-series data for 633 locations in 34 countries in the period 1972-2017. First, a hierarchical two-stage design was defined to estimate pooled temperature-mortality associations with nested levels of country/location random effects, borrowing information across multiple levels. Second, first-stage models were repeated in sub-periods and the estimates were pooled in multivariate longitudinal meta-analyses to examine location and country-specific variations in temperature-mortality risks. Third, varying effects by age were investigated through multivariate dose-response meta-regression models that allow the pooling of repeated estimates for different age groups across locations and countries. Fourth, effect modification by air conditioning in heat-related risks was assessed in multivariate longitudinal meta-regressions that account for both geographical and temporal differences in air conditioning prevalence.
Conclusion: These applications demonstrate the advantages of the extended two-stage framework and its software implementation for developing flexible and innovative study designs in environmental epidemiological studies.